Skip to main content
Glama
YunkaiLim
by YunkaiLim

Yunkai Desktop Bridge

A local-first Windows MCP bridge for guarded desktop perception and interaction.

This repository is a sanitized public-source export of the DesktopBridge used in the Yunkai project. It keeps the control boundaries that matter: semantic/read-only inspection first, bounded input actions, local runtime policy, deterministic verification, metadata-only action audit, and no arbitrary shell.

Highlights

  • Windows virtual-desktop and active-window screenshots

  • Microsoft UI Automation (UIA) semantic inspection

  • Win32 control fallback

  • guarded mouse and keyboard input

  • verified UIA click/text actions with expected-post predicates

  • bounded game-input mode behind a separate local permission tier

  • localhost-only optional vision fallback

  • local DPAPI-backed secret aliases; secret values are never accepted as remote MCP arguments

  • capability manifest, runtime permission state, routing hints, device snapshot, and action trace surfaces

Related MCP server: dsh-desktop-operator

Deliberate non-capabilities

The public MCP surface does not expose arbitrary shell/PowerShell/cmd execution, file deletion, software uninstall, shutdown/reboot, registry modification, unrestricted process kill, or remote secret upload.

Public export boundary

This repository does not contain private tunnel IDs, runtime credentials, owner-specific configuration, caches, backups, local verification evidence, or tunnel-client binaries. Machine-specific DevSpace routing code is intentionally excluded from this public bridge.

The shared contract modules required by the bridge are included under yunkai_shared/.

Requirements

  • Windows 10/11

  • Python 3.10+

  • Microsoft UI Automation (built into Windows)

  • Python dependencies in requirements.txt

Quick start

git clone https://github.com/YunkaiLim/yunkai-desktop-bridge.git
cd yunkai-desktop-bridge
py -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cd DesktopBridge
python server_stdio.py

For the local HTTP entry point:

cd DesktopBridge
python server.py

Optional localhost vision can be configured by copying DesktopBridge/desktopbridge_vision.example.json to desktopbridge_vision.json and selecting a compatible local vision model.

Tests

From DesktopBridge/:

python -m unittest

During public-release preparation on 2026-09-30, the selected public test set passed 104 tests, including stdio MCP discovery and screenshot/content checks.

A passing unit suite is not proof of live acceptance on every Windows build or MCP host. Perform target-machine acceptance separately.

Security model

Read SECURITY.md before enabling side-effecting tools. Keep runtime permissions local, preserve foreground-window and semantic-target guards, and do not remove verification simply to reduce latency.

License

Apache-2.0. See LICENSE and NOTICE.

Third-party components are not relicensed by this repository; see THIRD_PARTY_NOTICES.md.

This project is independent source code and is not an official OpenAI product.

Related MCP Connectors

  • Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.

  • Securely control computers you explicitly pair through files, terminals, processes, screenshots, desktop UI/input, clipboard, browser automation, diagnostics, and document tools.

  • Operate Linux, macOS and Windows from your LLM. Every action runs through an auditable allowlist.

  • Local Commander connects ChatGPT, Claude, Cursor and other MCP clients to Windows 11 computers you own. Your AI can read approved project files and Git state, run the checks and commands you authorize, and check local services. Your workspace at localcmd.app shows which computer ran each request and how it ended. Once you approve a computer and connect your AI, it can do what your Windows account can do; you can limit it to folders, block commands or switch it off at any time. Free on up to 5 computers; the Team pack adds 10 more for $15 a month.

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables safe Windows desktop automation and computer use through natural language, including window observation, UI Automation, and execution of verified actions like clicking, typing, and scrolling.
    2
    MIT
  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables secure, deny-by-default automation of Windows windows with per-action permissions, allowing observation, clicking, typing, and dragging on approved applications while verifying target identity and focus at runtime.
    Apache 2.0
  • A
    license
    Not graded
    quality
    C
    maintenance
    Enables AI agents to inspect and drive native Windows desktop applications (WinForms, WPF, WinUI 3, and WebView2 hybrids) through Microsoft UI Automation by capturing interactive control snapshots, then clicking, filling, and asserting on elements via ephemeral refs or semantic screen-map targets. Mutating actions are dry-run by default and gated behind explicit arming, allow-lists, and a kill switch for lawful, assisted QA testing.
    MIT